Loading ...
Sorry, an error occurred while loading the content.

Re: [extremeperl] refactoring databases

Expand Messages
  • Rob Nagler
    Hi Perrin, It s good to hear from you. ... Probably. It s certainly something we do. I ve included some thoughts below. ... This is like the code we have,
    Message 1 of 11 , Feb 18, 2005
    • 0 Attachment
      Hi Perrin,

      It's good to hear from you.

      > Does this sound like something you posted?

      Probably. It's certainly something we do. I've included some
      thoughts below.

      Matthew Albright writes:
      > It's easier to show some code that would work (as opposed to describing it in
      > words):

      This is like the code we have, but more complex. We don't actually
      store the schema change scripts in the database. We do have a
      db_upgrade_t (free with bOP), with a uniqueness constraint on
      the version field which is populated with the CVS $Revision$ field.

      I've appended one of our upgrade classes. We've done a bad job of
      sharing upgrade code across our applications. It's something on my
      todo list, and I finally turned it into a class a little while ago.
      The application I ripped this from doesn't do an export before the
      upgrade, because it is a huge database. All our other db-upgrade
      scripts do a full export before doing anything.

      > What this does is give you a table that grows over time that
      > indicates exactly which schema changes have been done to the
      > database (and when).

      We use CVS for this. The main problem is that many schema upgrades
      aren't schema upgrades at all. For example, we recently introduced an
      options database in bivio.com. Option ticker symbols are, well,
      stupid. Most brokers don't track historical option prices. However,
      we're an accounting service, and we have to -- Quicken doesn't, but
      there ya go. ;-) Anyways, we had to upgrade the names so they were
      temporally and spatially unique. We consider this a schema change,
      and programmed upgrade_1_options_tickers (see below for why this is
      named that way), even though there was no DDL associated with the
      change.

      > And if any part of the schema change fails, the transaction rolls
      > back and the row is not inserted into the schema_change table, so
      > you can try again after you fix your code.

      This doesn't work with Oracle as mentioned by Johan below.

      > So, you can ship out this update_db.pl file to your customers and be
      > confident that no matter how old their database schema is, it will
      > be up to date after this file is run.

      This is important, and we do this. You test this script to death, and
      then you still get it wrong sometimes. :-(

      > And like you said, every so often you do a check of all your
      > customers, and figure out the oldest version of code that is out in
      > the wild that has a chance of upgrading in the future.

      This is simply not an option we offer our customers. We could do
      this, I guess, but I don't believe in multiple versions in the field
      running wild. At most two: test and prod.

      > If anyone is interested, I can share some additional thoughts on
      > data migration landmines to watch out for...

      Please do.

      Johan Lindstrom writes:
      > In my experience, doing this automatically is kind of dangerous.

      Please elaborate why this type of automation is dangerous? We've had
      about 200+ schema upgrades to bivio.com, 22 upgrades to colosla.org,
      and 26 for myBallot at ieee.org, which is still in beta. I can't
      release stats for other sites, but the numbers are similar. I can get
      these stats in a few seconds from CVS. It's useful data.

      > The application of changes in the database is done manually also, mainly
      > because we can't at _all_ times be certain that someone didn't fiddle with
      > something during an emergency fixup.

      If someone fiddles, and they don't check-in their changes, they burn
      for fiddling. Everything is checked in to CVS. No exceptions.

      > And in many databases (Sybase, Oracle), DDL can't be done in a
      > transaction and most of the time a schema change involves both a
      > structure change and some kind of data transformation to support
      > this. Sometimes it requires two alters with a data transformation in
      > between.

      Yup, all the more reason to do this programmatically, and test the
      heck out of it. Here's our normal procedure:

      1) Develop the procedure, update the code, test on your dev database
      2) Bring over a copy of the production database (if possible), and
      test on your dev database.
      3) Release, and run on test system.
      4) Release, and run on production.

      If we are being particularly paranoid, we ask everybody to upgrade
      their dev databases before going to test.

      Also, we do parallel development. There's upgrade_1_*, upgrade_2_*,
      etc. You have to have a system for this so people can apply updates
      incrementally to their dev and the test databases. Once released to
      production, we apply all updates with one command. If one of the
      commands, blows, you can initiate individual methods by hand.
      (Bivio::ShellUtil makes this easy for us, but that's another story.)

      > On a general note, having developers change schemas with no sense of the
      > volumes of data in the tables can be very bad.

      Having developers with no sense of volumes of data is very bad. We
      all learn this by making that mistake a few times. It's also bad to
      have the developers test on faster development machines without
      knowing there's going to be a slow down when it runs on slower
      production machines.

      > Sometimes one can prepare the migration and just do the most critical part
      > during the downtime.

      We try to do this, too.

      > Of course this depends on the size of the shop and how well the sysadmins
      > communicate with the developers and how well versed the developers are with
      > db-related concepts.

      This is why it is extremely important for developers to be dbas and
      sysadmins, too. They have to know that problems can occur at all
      levels in the system. Indeed, most hard problems occur between large
      subsystems, such as, DBMS and application or application and MTA.

      At bivio, programmers are responsible for the end-to-end service we
      offer our customers: from conception to support with all the
      customer-irrelevant bits inbetween, like programming and sysadmin. We
      often are responsible for marketing and sales, too. It's humbling to
      try to sell your beautifully crafted DB-backed application to someone
      who can't even spell DB, and doesn't want to. They just want to have
      their books balance at the end of the year, and they want their tax
      return to print out (even when Apple's Preview for PDF doesn't handle
      FDF, and Acrobat 7.0 is not downloadable to Windows 98, and ... -- ok,
      it's tax season, and we'll feeling the pain right now. :-)

      Some other things we've learned, not all of which have been put in
      place. Some of these suggestions will no doubt generate shrieks from
      DBAs. :-) Most of these will help greatly if you like the XP
      development model.

      * Universally unique sequence ids. This is very useful for debugging,
      and scalability.

      * Generate the entire schema from your application models. We still
      don't do this, and it means we have to maintain three files: *.sql,
      perl models, and DbUpgrade.pm.

      * Release early, release often. This is the best way to ensure schema
      changes remain small, and lead to minimal system disruption, when
      something goes wrong.

      * With big databases, buy a new machine when you upgrade the DBMS
      software. Use Oracle's standby database feature to minimize
      downtime. Don't do ANY changes (schema or app) when you do this.

      * Automate the things you do most often. If you release often (did
      I mention this?), make sure it takes about 1 or 2 minutes to cut
      a release, and 1 or 2 minutes to install it.

      * Use your app to validate all data before it hits the database except
      unqiueness constraints. Use the db constraints as a backup for you
      software. Db consistency is the most important issue you face. If
      you see a

      * Disk is cheap. Don't worry about how the database stores things
      until it becomes a problem. LOBs are still a problem with DBMSes,
      and we have learned that lesson the hard way.

      * Add indexes wherever you can, but don't preoptimize. Solve
      performance problems as you encounter them. Don't stress about
      the physical structure. It's probably not an issue at all.

      * The DBMS is your bottleneck when it comes to scalability and
      maintainability. Don't use stored procedures. Don't use special
      features. I am still kicking myself over the use of Oracle's CONNECT
      BY. It was a stupid preoptimization that was not necessary. As it turns
      out, Oracle 8i's implementation of self-related tables is horribly
      buggy. We are slowly in the process of removing this
      preoptimization to no performance detriment, and probably to the
      contrary, a general speed up.

      * Automate all tests and run them every day, automatically. Tests
      which aren't automated don't get run.

      This is our experience. Here's the performance:

      http://www.bivio.com/club_index/accounting/investments

      The database server is a 450mhz quad processor with 2GB of memory and
      10K disks running Oracle 8i. The schema that produces this table is
      fully normalized. You are seeing the sum of all transactions in
      tables that hold millions of transactions for all of our customers.
      The prices table has hundreds of millions of records. Nothing is
      cached except what Oracle caches.

      Rob
      ----------------------------------------------------------------
      # Copyright (c) 2005 bivio Software, Inc. All Rights Reserved.
      package Bivio::SQL::DbUpgrade;
      use strict;
      $Bivio::SQL::DbUpgrade::VERSION = sprintf('%d.%02d', q$Revision: 2.22 $ =~ /\d+/g);
      $_ = $Bivio::SQL::DbUpgrade::VERSION;

      use Bivio::ShellUtil;
      @Bivio::SQL::DbUpgrade::ISA = ('Bivio::ShellUtil');

      sub USAGE {
      return <<'EOF';
      usage: s-db-upgrade [options] command [args...]
      commands:
      do_all -- run upgrade method
      upgrade_1_<name> -- first upgrade method
      EOF
      }

      use Bivio::IO::File;
      use Bivio::IO::Ref;
      use Bivio::Type::Date;
      use Bivio::Type::DateInterval;
      use Bivio::Type::DateTime;

      sub do_all {
      my($self) = @_;
      local($|) = 1;
      $self->initialize_ui;
      my($req) = $self->get_request;
      my($upgrade) = Bivio::Biz::Model->new($req, 'DbUpgrade');
      my($v) = $Bivio::SQL::DbUpgrade::VERSION;
      $self->usage("$v already ran")
      if $upgrade->unauth_load(version => $v);
      foreach my $method (_methods()) {
      $self->print("Calling: $method\n");
      $self->$method();
      }
      # creates a completion entry in db to prevent re-running
      $upgrade->create({
      version => $v,
      run_date_time => Bivio::Type::DateTime->now
      });
      return "*** done ***\n";
      }

      sub upgrade_1_some_change {
      my($self) = @_;
      "do something here";
      return;
      }


      sub _add_column {
      my($table, $column, $type, $value) = @_;
      print "Adding $table.$column\n";
      _do_sql("
      ALTER TABLE $table
      ADD $column $type
      /
      ");
      return unless defined($value);

      # Insert default values
      print "Inserting default into $table.$column\n";
      _do_sql("
      update $table set $column = $value
      /");

      return;
      }

      sub _backup_model {
      my($model, $order_by, $req) = @_;
      my($rows) = [];
      my($m) = Bivio::Biz::Model->new($req, $model);
      for ($m->unauth_iterate_start($order_by); $m->iterate_next_and_load;) {
      push(@$rows, $m->get_shallow_copy);
      }
      $m->iterate_end;
      my($f) = "$model-" . Bivio::Type::DateTime->local_now_as_file_name . '.PL';
      Bivio::IO::File->write($f, Bivio::IO::Ref->to_string($rows));
      print("Backed up $f\n");
      return $rows;
      }

      sub _do_sql {
      my($sql) = @_;
      my($statement);
      my($s) = '';
      foreach my $line (split(/\n/, $sql)) {
      # Skip comments and blanks
      next if $line =~ /^\s*--|^\s*$/;

      # Execute statement if '/' found
      if ($line =~ /^\s*\/\s*$/) {
      $statement = Bivio::SQL::Connection->execute($s);
      $s = '';
      next;
      }

      # Build up statement
      $s .= $line."\n";
      }
      die("$s: left over statement") if $s;
      return $statement;
      }

      sub _drop_column {
      my($table, $column) = @_;
      print "Dropping $table.$column\n";
      _do_sql("
      ALTER TABLE $table
      DROP COLUMN $column
      /
      ");
      return;
      }

      sub _drop_constraints {
      my($table, $column) = map {uc($_)} @_;
      # Find all constraints on the table
      my($statement) = _do_sql(<<"EOF");
      SELECT user_cons_columns.constraint_name
      FROM user_cons_columns, user_constraints
      WHERE user_cons_columns.column_name = '$column'
      AND user_cons_columns.table_name = '$table'
      AND user_constraints.constraint_name
      = user_cons_columns.constraint_name
      /
      EOF
      my(@constraints);
      while (my($name) = $statement->fetchrow_array) {
      push(@constraints, $name);
      }
      $statement->finish;

      # Drop the constraints
      foreach my $c (@constraints) {
      _do_sql("ALTER TABLE $table drop constraint $c
      /");
      }
      return int(@constraints);
      }

      sub _init_realm_role {
      my($rru, $perms) = @_;
      $perms =~ s/\\//g;
      $perms =~ s/^.*b-realm-role.* edit /!/mg;
      foreach my $edit (split(/!/, $perms)) {
      next unless $edit;
      $rru->edit(split(' ', $edit));
      }
      return;
      }

      sub _methods {
      {
      no strict;
      local($^W) = 0;;
      local(*stab) = eval('*' . __PACKAGE__ . '::');
      return grep(
      $_ =~ /^upgrade_\d+_/ && __PACKAGE__->can($_),
      sort(keys(%stab)),
      );
      use strict;
      }
      }

      sub _modify_column {
      my($table, $column, $type) = @_;
      _do_sql("
      ALTER TABLE $table MODIFY $column $type
      /");
      return;
      }

      sub _sqlplus {
      my($db_login, $actions) = @_;
      my($res) = Bivio::ShellUtil::piped_exec(undef, 'su - oracle -c sqlplus',
      <<"EOF");
      $db_login
      set linesize 10000
      set pagesize 10000
      whenever sqlerror exit sql.sqlcode rollback
      whenever oserror exit sql.sqlcode rollback
      $actions;
      disconnect
      exit
      EOF
      return $res;
      }

      1;
    • Johan Lindstrom
      ... This is just my experience at $work. But you should know that we re not very mature on process at all, and currently there is no way to guarantee that
      Message 2 of 11 , Feb 18, 2005
      • 0 Attachment
        At 18:37 2005-02-18, Rob Nagler wrote:
        > > In my experience, doing this automatically is kind of dangerous.
        >
        >Please elaborate why this type of automation is dangerous? We've had

        This is just my experience at $work. But you should know that we're not
        very mature on process at all, and currently there is no way to guarantee
        that there have been no fiddling. That's the current goal, to ensure we're
        at a known state before each upgrade.

        It's not quite best practice, but I guess a lot of shops are in a somewhat
        similar situation, especially if they have a tradition of strong operations
        departments with strong dba:s.


        > > Sometimes one can prepare the migration and just do the most critical part
        > > during the downtime.
        >
        >We try to do this, too.

        Do you automate these steps also?


        > > Of course this depends on the size of the shop and how well the sysadmins
        > > communicate with the developers and how well versed the developers are
        > with
        > > db-related concepts.
        >
        >This is why it is extremely important for developers to be dbas and
        >sysadmins, too. They have to know that problems can occur at all

        Agree 100%. It's a simple fact that all developers need to know their
        tools. And a db is a tool being used similar to a compiler. It's not a
        black box where there's... uh, data or something.


        >* Universally unique sequence ids. This is very useful for debugging,
        > and scalability.

        Does that mean you don't share identities between tables at all? Just one
        global sequence (or whatever)?


        /J

        -------- ------ ---- --- -- -- -- - - - - -
        Johan Lindström Sourcerer @ Boss Casinos johanl AT DarSerMan.com

        Latest bookmark: "TCP Connection Passing"
        http://tcpcp.sourceforge.net/
        dmoz: /Computers/Programming/Languages/JavaScript/ 12
      • Matthew Albright
        ... Yeah, we re on postgresql. ... We are actually shipping a software product, and we make the new releases available to our customers, but they don t have to
        Message 3 of 11 , Feb 18, 2005
        • 0 Attachment
          --- Rob Nagler <nagler@...> wrote:
          > This doesn't work with Oracle as mentioned by Johan below.

          Yeah, we're on postgresql.

          > This is simply not an option we offer our customers. We could do
          > this, I guess, but I don't believe in multiple versions in the field
          > running wild. At most two: test and prod.

          We are actually shipping a software product, and we make the new releases
          available to our customers, but they don't have to upgrade every time we
          release. We're an XP shop, so we release often, which means we have a wide
          variety of versions in the field.

          > This is why it is extremely important for developers to be dbas and
          > sysadmins, too.

          I couldn't agree more with this. When whoever it was started talking about
          DBA's changing the database structure, my eyes glazed over, because I've never
          been in an organization that had a DBA. We have people who are stronger and
          weaker in certain areas, but really, anyone can modify anything... that's what
          pair programming is for.

          > > If anyone is interested, I can share some additional thoughts on
          > > data migration landmines to watch out for...
          >
          > Please do.

          Like most relational database-backed applications, we've coded up a
          object-relational mapping layer that follows the "Active Record" pattern
          (http://www.martinfowler.com/eaaCatalog/activeRecord.html). The problems come
          in when you call too much of that mapping code from your update_db.pl script.
          At the time the update_db.pl code is run, it's running against the newest
          production code. If your migration code calls methods in the "real" code, that
          code may change, and your migration code might cease to work in the future.

          To avoid this rather subtle problem (which I suspect I've not explained very
          lucidly), we do 2 things:

          - Avoid the use of "real" code in the migration code, even if it means
          duplication
          - Test to make sure it is possible to upgrade from the oldest customer
          version to the new version

          Sometimes, the use of real code is unavoidable, and we've had to reorder things
          in the update_db.pl file... that's something to be avoided.

          matt
        • Adrian Howard
          On 19 Feb 2005, at 00:50, Matthew Albright wrote: [snip] ... [snip] ... Amen. ... Just to clarify - I was talking about a /client/ DBA who tweaked the schema
          Message 4 of 11 , Feb 19, 2005
          • 0 Attachment
            On 19 Feb 2005, at 00:50, Matthew Albright wrote:
            [snip]
            > --- Rob Nagler <nagler@...> wrote:
            [snip]
            >> This is why it is extremely important for developers to be dbas and
            >> sysadmins, too.
            >
            > I couldn't agree more with this.

            Amen.

            > When whoever it was started talking about
            > DBA's changing the database structure, my eyes glazed over, because
            > I've never
            > been in an organization that had a DBA.

            Just to clarify - I was talking about a /client/ DBA who tweaked the
            schema of software that we shipped to them - not the DBA of the
            development team.

            Not that we had anybody with that title.

            (although we did have an Oracle goddess who got paired with a heck of a
            lot when it came to the DB code :-)

            Yes this was (from one point of view) the clients "fault" - but if
            you're going to ship them automated scripts to update there database
            having an automated double-check that you're really updating the right
            schema is a nice thing.

            Adrian
          • Rob Nagler
            ... Yes, we try our best to automate everything. If we think there is going to be multiple upgrades, one of which is critical, we ll organize these in
            Message 5 of 11 , Feb 19, 2005
            • 0 Attachment
              Johan Lindstrom writes:
              > > > Sometimes one can prepare the migration and just do the most critical part
              > > > during the downtime.
              > >
              > >We try to do this, too.
              >
              > Do you automate these steps also?

              Yes, we try our best to automate everything. If we think there is
              going to be multiple upgrades, one of which is critical, we'll
              organize these in separate releases. If this isn't possible, we'll
              make it a two-part upgrade. That script I included in a previous
              message doesn't update the db_upgrade_t if you just call one of the
              methods.

              > Does that mean you don't share identities between tables at all? Just one
              > global sequence (or whatever)?

              Not exactly. The following describes it in detail:

              http://petshop.bivio.biz/src?s=Bivio::Type::PrimaryId

              This the base class of all serial keys. A PrimaryId is a NUMERIC(18)
              (but could be larger) which looks like:

              18 5 3 2 0
              +------------------+-------------+----------+-------+
              | Monotonically | Site Id or | Mistake | Type |
              | Increasing Seq | Other info | Bit | |
              +------------------+-------------+----------+-------+

              The numbers above are digit index. Type is a number that uniquely
              identifies the key type. This divides the name space. The Site Id
              allows you to divide the name space across independent databases. The
              Mistake Bit allows you to make mistakes. :-) We used it in one case
              to transform one key into another without changing the type It saved
              a lookup, and was related to finding the start of a tree for Oracle's
              CONNECT BY. Now they we are dumping CONNECT BY, we have no need for
              it. But that's another discussion.

              Here's how we initialize the sequences:

              CREATE sequence user_s
              MINVALUE 100001
              CACHE 1 INCREMENT BY 100000
              /

              CREATE SEQUENCE club_s
              MINVALUE 100002
              CACHE 1 INCREMENT BY 100000
              /

              CREATE SEQUENCE ec_payment_s
              MINVALUE 100015
              CACHE 1 INCREMENT BY 100000
              /

              CREATE sequence bulletin_s
              MINVALUE 100016
              CACHE 1 INCREMENT BY 100000
              /

              As you can see the spaces are unique, but overlapping. Some people
              might complain it junks up the indexes, but we haven't seen that to be
              a problem. Once the tables get big enough, it really doesn't matter.

              Rob
            • Rob Nagler
              ... Perhaps you should pay them to upgrade. :-) Yes, you ve got a logistical problem, and it sounds like you have a good solution for it. ... Interesting you
              Message 6 of 11 , Feb 19, 2005
              • 0 Attachment
                Matthew Albright writes:
                > We are actually shipping a software product, and we make the new releases
                > available to our customers, but they don't have to upgrade every time we
                > release. We're an XP shop, so we release often, which means we have a wide
                > variety of versions in the field.

                Perhaps you should pay them to upgrade. :-) Yes, you've got a
                logistical problem, and it sounds like you have a good solution for
                it.

                > been in an organization that had a DBA. We have people who are stronger and
                > weaker in certain areas, but really, anyone can modify anything... that's what
                > pair programming is for.

                Interesting you should mention that. I've been a bit turned off to
                "general pair programming" lately. We pair for the same reason. One
                of the things about pair programming that I don't like is that when
                people are alone, they learn from their mistakes at their own speed.
                If they are with another programmer, they not only get the problem
                solved more quickly, but they also don't necessarily assimilate the
                thought process and tools the other programmer brins to the table.
                Experience caused by thousands of wrong paths is a great guide.
                Finding the path right away isn't always the best way. It's certainly
                my experience with kids that they learn much better when they burn
                their hands than when you tell them the iron is hot.

                > Like most relational database-backed applications, we've coded up a
                > object-relational mapping layer that follows the "Active Record" pattern
                > (http://www.martinfowler.com/eaaCatalog/activeRecord.html). The problems come
                > in when you call too much of that mapping code from your update_db.pl script.
                > At the time the update_db.pl code is run, it's running against the newest
                > production code. If your migration code calls methods in the "real" code, that
                > code may change, and your migration code might cease to work in the future.

                Excellent point. We see this sometimes when trying to do a current
                upgrade. I would hate to solve this problem for N versions in the
                past.

                > To avoid this rather subtle problem (which I suspect I've not explained very
                > lucidly), we do 2 things:
                >
                > - Avoid the use of "real" code in the migration code, even if it means
                > duplication
                > - Test to make sure it is possible to upgrade from the oldest customer
                > version to the new version
                >
                > Sometimes, the use of real code is unavoidable, and we've had to reorder things
                > in the update_db.pl file... that's something to be avoided.

                Great solution. Seems like you have to do a lot of feature-based
                tests. For example, you might add a feature, fix it, then remove it
                (because it wasn't the right solution in the first place). Those
                three upgrades are a no-op after the "remove it" upgrade.

                One issue I don't see how you get around is the collection of new
                types of data. You can never assume that a new data type of data
                exists in the database. This means you probably can't retire
                "deprecated" code that works around the non-existence of the new type
                of data. This may create lots and lots of paths to test. Ugh...

                Rob
              • Rob Nagler
                ... Yikes. Remember there s a clear separation of concerns between the customer and the programmer. Be sure to yell at your customer. ;-) ... Yup. I wonder
                Message 7 of 11 , Feb 19, 2005
                • 0 Attachment
                  Adrian Howard writes:
                  > Just to clarify - I was talking about a /client/ DBA who tweaked the
                  > schema of software that we shipped to them - not the DBA of the
                  > development team.

                  Yikes. Remember there's a clear separation of concerns between the
                  customer and the programmer. Be sure to yell at your customer. ;-)

                  > Yes this was (from one point of view) the clients "fault" - but if
                  > you're going to ship them automated scripts to update there database
                  > having an automated double-check that you're really updating the right
                  > schema is a nice thing.

                  Yup. I wonder though if you don't also have to protect the data?
                  Usually apps trust the data that comes from the database. In your
                  case, that wouldn't be the case. Do you have extra validation to make
                  sure the data hasn't been tampered with?

                  Rob
                Your message has been successfully submitted and would be delivered to recipients shortly.